Add rustc option to output LLVM optimization remarks to YAML files

Author: Kobzol

compiler/rustc_codegen_llvm/src/back/lto.rs

@@ -1,4 +1,4 @@
-use crate::back::write::{self, save_temp_bitcode, DiagnosticHandlers};
+use crate::back::write::{self, save_temp_bitcode, CodegenDiagnosticsSection, DiagnosticHandlers};
 use crate::errors::{
     DynamicLinkingWithLTO, LlvmError, LtoBitcodeFromRlib, LtoDisallowed, LtoDylib,
 };
@@ -302,7 +302,13 @@ fn fat_lto(
         // The linking steps below may produce errors and diagnostics within LLVM
         // which we'd like to handle and print, so set up our diagnostic handlers
         // (which get unregistered when they go out of scope below).
-        let _handler = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
+        let _handler = DiagnosticHandlers::new(
+            cgcx,
+            diag_handler,
+            llcx,
+            &module,
+            CodegenDiagnosticsSection::LTO,
+        );
 
         // For all other modules we codegened we'll need to link them into our own
         // bitcode. All modules were codegened in their own LLVM context, however,

compiler/rustc_codegen_llvm/src/back/write.rs

@@ -268,6 +268,13 @@ pub(crate) fn save_temp_bitcode(
     }
 }
 
+/// In what context is a dignostic handler being attached to a codegen unit?
+pub enum CodegenDiagnosticsSection {
+    Codegen,
+    Opt,
+    LTO,
+}
+
 pub struct DiagnosticHandlers<'a> {
     data: *mut (&'a CodegenContext<LlvmCodegenBackend>, &'a Handler),
     llcx: &'a llvm::Context,
@@ -279,6 +286,8 @@ impl<'a> DiagnosticHandlers<'a> {
         cgcx: &'a CodegenContext<LlvmCodegenBackend>,
         handler: &'a Handler,
         llcx: &'a llvm::Context,
+        module: &ModuleCodegen<ModuleLlvm>,
+        section: CodegenDiagnosticsSection,
     ) -> Self {
         let remark_passes_all: bool;
         let remark_passes: Vec<CString>;
@@ -295,6 +304,21 @@ impl<'a> DiagnosticHandlers<'a> {
         };
         let remark_passes: Vec<*const c_char> =
             remark_passes.iter().map(|name: &CString| name.as_ptr()).collect();
+        let remark_file = cgcx
+            .remark_dir
+            .as_ref()
+            .and_then(|dir| dir.to_str())
+            // Use the .opt.yaml file pattern, which is supported by LLVM's opt-viewer.
+            .map(|dir| {
+                let section = match section {
+                    CodegenDiagnosticsSection::Codegen => "codegen",
+                    CodegenDiagnosticsSection::Opt => "opt",
+                    CodegenDiagnosticsSection::LTO => "lto",
+                };
+                format!("{dir}/{}.{section}.opt.yaml", module.name)
+            })
+            .map(|dir| CString::new(dir).unwrap());
+
         let data = Box::into_raw(Box::new((cgcx, handler)));
         unsafe {
             let old_handler = llvm::LLVMRustContextGetDiagnosticHandler(llcx);
@@ -305,6 +329,9 @@ impl<'a> DiagnosticHandlers<'a> {
                 remark_passes_all,
                 remark_passes.as_ptr(),
                 remark_passes.len(),
+                // The `as_ref()` is important here, otherwise the `CString` will be dropped
+                // too soon!
+                remark_file.as_ref().map(|dir| dir.as_ptr()).unwrap_or(std::ptr::null()),
             );
             DiagnosticHandlers { data, llcx, old_handler }
         }
@@ -523,7 +550,8 @@ pub(crate) unsafe fn optimize(
 
     let llmod = module.module_llvm.llmod();
     let llcx = &*module.module_llvm.llcx;
-    let _handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
+    let _handlers =
+        DiagnosticHandlers::new(cgcx, diag_handler, llcx, module, CodegenDiagnosticsSection::Opt);
 
     let module_name = module.name.clone();
     let module_name = Some(&module_name[..]);
@@ -582,7 +610,13 @@ pub(crate) unsafe fn codegen(
         let tm = &*module.module_llvm.tm;
         let module_name = module.name.clone();
         let module_name = Some(&module_name[..]);
-        let handlers = DiagnosticHandlers::new(cgcx, diag_handler, llcx);
+        let _handlers = DiagnosticHandlers::new(
+            cgcx,
+            diag_handler,
+            llcx,
+            &module,
+            CodegenDiagnosticsSection::Codegen,
+        );
 
         if cgcx.msvc_imps_needed {
             create_msvc_imps(cgcx, llcx, llmod);
@@ -775,7 +809,6 @@ pub(crate) unsafe fn codegen(
         }
 
         record_llvm_cgu_instructions_stats(&cgcx.prof, llmod);
-        drop(handlers);
     }
 
     // `.dwo` files are only emitted if:

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -2512,6 +2512,7 @@ extern "C" {
         remark_all_passes: bool,
         remark_passes: *const *const c_char,
         remark_passes_len: usize,
+        remark_file: *const c_char,
     );
 
     #[allow(improper_ctypes)]

compiler/rustc_codegen_ssa/src/back/write.rs

@@ -345,6 +345,9 @@ pub struct CodegenContext<B: WriteBackendMethods> {
     pub diag_emitter: SharedEmitter,
     /// LLVM optimizations for which we want to print remarks.
     pub remark: Passes,
+    /// Directory into which should the LLVM optimization remarks be written.
+    /// If `None`, they will be written to stderr.
+    pub remark_dir: Option<PathBuf>,
     /// Worker thread number
     pub worker: usize,
     /// The incremental compilation session directory, or None if we are not
@@ -1041,6 +1044,16 @@ fn start_executing_work<B: ExtraBackendMethods>(
             tcx.backend_optimization_level(())
         };
     let backend_features = tcx.global_backend_features(());
+
+    let remark_dir = if let Some(ref dir) = sess.opts.unstable_opts.remark_dir {
+        // Should this be here?
+        // Can this conflict with a parallel attempt to create this directory?
+        fs::create_dir_all(dir).expect("Cannot create remark directory");
+        Some(dir.canonicalize().expect("Cannot canonicalize remark directory"))
+    } else {
+        None
+    };
+
     let cgcx = CodegenContext::<B> {
         crate_types: sess.crate_types().to_vec(),
         each_linked_rlib_for_lto,
@@ -1052,6 +1065,7 @@ fn start_executing_work<B: ExtraBackendMethods>(
         prof: sess.prof.clone(),
         exported_symbols,
         remark: sess.opts.cg.remark.clone(),
+        remark_dir,
         worker: 0,
         incr_comp_session_dir: sess.incr_comp_session_dir_opt().map(|r| r.clone()),
         cgu_reuse_tracker: sess.cgu_reuse_tracker.clone(),

compiler/rustc_llvm/llvm-wrapper/RustWrapper.cpp

@@ -7,7 +7,12 @@
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/Intrinsics.h"
 #include "llvm/IR/IntrinsicsARM.h"
+#include "llvm/IR/LLVMRemarkStreamer.h"
 #include "llvm/IR/Mangler.h"
+#include "llvm/Remarks/RemarkStreamer.h"
+#include "llvm/Remarks/RemarkSerializer.h"
+#include "llvm/Remarks/RemarkFormat.h"
+#include "llvm/Support/ToolOutputFile.h"
 #if LLVM_VERSION_GE(16, 0)
 #include "llvm/Support/ModRef.h"
 #endif
@@ -1855,23 +1860,41 @@ using LLVMDiagnosticHandlerTy = DiagnosticHandler::DiagnosticHandlerTy;
 // When RemarkAllPasses is true, remarks are enabled for all passes. Otherwise
 // the RemarkPasses array specifies individual passes for which remarks will be
 // enabled.
+//
+// If RemarkFilePath is not NULL, optimization remarks will be streamed directly into this file,
+// bypassing the diagnostics handler.
 extern "C" void LLVMRustContextConfigureDiagnosticHandler(
     LLVMContextRef C, LLVMDiagnosticHandlerTy DiagnosticHandlerCallback,
     void *DiagnosticHandlerContext, bool RemarkAllPasses,
-    const char * const * RemarkPasses, size_t RemarkPassesLen) {
+    const char * const * RemarkPasses, size_t RemarkPassesLen,
+    const char * RemarkFilePath
+) {
 
   class RustDiagnosticHandler final : public DiagnosticHandler {
   public:
-    RustDiagnosticHandler(LLVMDiagnosticHandlerTy DiagnosticHandlerCallback,
-                          void *DiagnosticHandlerContext,
-                          bool RemarkAllPasses,
-                          std::vector<std::string> RemarkPasses)
+    RustDiagnosticHandler(
+      LLVMDiagnosticHandlerTy DiagnosticHandlerCallback,
+      void *DiagnosticHandlerContext,
+      bool RemarkAllPasses,
+      std::vector<std::string> RemarkPasses,
+      std::unique_ptr<llvm::remarks::RemarkStreamer> RemarkStreamer,
+      std::unique_ptr<ToolOutputFile> RemarksFile
+    )
         : DiagnosticHandlerCallback(DiagnosticHandlerCallback),
           DiagnosticHandlerContext(DiagnosticHandlerContext),
           RemarkAllPasses(RemarkAllPasses),
-          RemarkPasses(RemarkPasses) {}
+          RemarkPasses(std::move(RemarkPasses)),
+          RemarkStreamer(std::move(RemarkStreamer)),
+          RemarksFile(std::move(RemarksFile)) {}
 
     virtual bool handleDiagnostics(const DiagnosticInfo &DI) override {
+      if (this->RemarkStreamer) {
+        if (auto *OptDiagBase = dyn_cast<DiagnosticInfoOptimizationBase>(&DI)) {
+          LLVMRemarkStreamer Streamer(*this->RemarkStreamer);
+          Streamer.emit(*OptDiagBase);
+          return true;
+        }
+      }
       if (DiagnosticHandlerCallback) {
         DiagnosticHandlerCallback(DI, DiagnosticHandlerContext);
         return true;
@@ -1912,14 +1935,51 @@ extern "C" void LLVMRustContextConfigureDiagnosticHandler(
 
     bool RemarkAllPasses = false;
     std::vector<std::string> RemarkPasses;
+    std::unique_ptr<llvm::remarks::RemarkStreamer> RemarkStreamer;
+    std::unique_ptr<ToolOutputFile> RemarksFile;
   };
 
   std::vector<std::string> Passes;
   for (size_t I = 0; I != RemarkPassesLen; ++I)
+  {
     Passes.push_back(RemarkPasses[I]);
+  }
+
+  // We need to hold onto both the streamer and the opened file
+  std::unique_ptr<llvm::remarks::RemarkStreamer> RemarkStreamer;
+  std::unique_ptr<ToolOutputFile> RemarkFile;
+
+  if (RemarkFilePath != nullptr) {
+    std::error_code EC;
+    RemarkFile = std::make_unique<ToolOutputFile>(
+      RemarkFilePath,
+      EC,
+      llvm::sys::fs::OF_TextWithCRLF
+    );
+    // Do not delete the file after we gather remarks
+    RemarkFile->keep();
+
+    auto RemarkSerializer = remarks::createRemarkSerializer(
+      llvm::remarks::Format::YAML,
+      remarks::SerializerMode::Separate,
+      RemarkFile->os()
+    );
+    if (Error E = RemarkSerializer.takeError())
+    {
+      // Is this OK?
+      report_fatal_error("Cannot create remark serializer");
+    }
+    RemarkStreamer = std::make_unique<llvm::remarks::RemarkStreamer>(std::move(*RemarkSerializer));
+  }
 
   unwrap(C)->setDiagnosticHandler(std::make_unique<RustDiagnosticHandler>(
-      DiagnosticHandlerCallback, DiagnosticHandlerContext, RemarkAllPasses, Passes));
+    DiagnosticHandlerCallback,
+    DiagnosticHandlerContext,
+    RemarkAllPasses,
+    Passes,
+    std::move(RemarkStreamer),
+    std::move(RemarkFile)
+  ));
 }
 
 extern "C" void LLVMRustGetMangledName(LLVMValueRef V, RustStringRef Str) {

compiler/rustc_session/src/config.rs

@@ -2583,6 +2583,10 @@ pub fn build_session_options(
         handler.early_warn("-C remark requires \"-C debuginfo=n\" to show source locations");
     }
 
+    if cg.remark.is_empty() && unstable_opts.remark_dir.is_some() {
+        handler.early_warn("using -C remark-dir without enabling remarks using e.g. -C remark=all");
+    }
+
     let externs = parse_externs(handler, matches, &unstable_opts);
 
     let crate_name = matches.opt_str("crate-name");

compiler/rustc_session/src/options.rs

@@ -1314,7 +1314,7 @@ options! {
         "control generation of position-independent code (PIC) \
         (`rustc --print relocation-models` for details)"),
     remark: Passes = (Passes::Some(Vec::new()), parse_passes, [UNTRACKED],
-        "print remarks for these optimization passes (space separated, or \"all\")"),
+        "output remarks for these optimization passes (space separated, or \"all\")"),
     rpath: bool = (false, parse_bool, [UNTRACKED],
         "set rpath values in libs/exes (default: no)"),
     save_temps: bool = (false, parse_bool, [UNTRACKED],
@@ -1659,6 +1659,9 @@ options! {
         "choose which RELRO level to use"),
     remap_cwd_prefix: Option<PathBuf> = (None, parse_opt_pathbuf, [TRACKED],
         "remap paths under the current working directory to this path prefix"),
+    remark_dir: Option<PathBuf> = (None, parse_opt_pathbuf, [UNTRACKED],
+        "directory into which to write optimization remarks (if not specified, they will be \
+written to standard error output)"),
     report_delayed_bugs: bool = (false, parse_bool, [TRACKED],
         "immediately print bugs registered with `delay_span_bug` (default: no)"),
     sanitizer: SanitizerSet = (SanitizerSet::empty(), parse_sanitizers, [TRACKED],